CN110114195B

CN110114195B - Action transfer device, action transfer method, and non-transitory computer-readable medium storing action transfer program

Info

Publication number: CN110114195B
Application number: CN201780068803.2A
Authority: CN
Inventors: 榎本昇平; 长谷川修
Original assignee: Soinn Inc
Current assignee: Soinn Inc
Priority date: 2016-11-07
Filing date: 2017-11-01
Publication date: 2022-05-17
Anticipated expiration: 2037-11-01
Also published as: US11207773B2; JP6518982B2; US20190283246A1; EP3536467A4; EP3536467A1; EP3536467B1; CN110114195A; JPWO2018084164A1; WO2018084164A1

Abstract

A transfer source motion information acquisition unit acquires first motion information of the transfer source robot. The transfer target motion information acquisition unit acquires second motion information of the transfer source robot. A correction unit generates third motion information of the transfer target robot by correcting the motion information of the transfer source robot using the second motion information and according to a predetermined formula. The number of data included in the second operation information of the transfer target robot is smaller than the number of data included in the first operation information of the transfer source robot. The first to third motion information includes a set of data indicating one or more robot joint values and a set of data indicating coordinate values of a specific part of the robot.

Description

Action transfer device, action transfer method, and non-transitory computer-readable medium storing action transfer program

Technical Field

The present invention relates to a motion transfer device, a motion transfer method, and a motion transfer program. For example, the present invention relates to a technique for transferring an action from a transfer source robot by using a small number of learning samples acquired by a transfer target robot.

Background

Currently, the use of robots has expanded in human society, including homes. At present, an intelligent development mechanism for a robot is still under development, and thus actions that the robot can perform and things that the robot can understand are limited. On the other hand, various types of robots having different physical characteristics have been developed. Assuming that robots having the above limitations are introduced into a standard home, it is inefficient to have each of these robots learn actions, etc. independently.

Therefore, a method for sharing knowledge, particularly motion information, learned by each robot among other robots is required. A technique for obtaining an action by sharing action information in a robot is called "action transfer". In other words, motion transfer is a technique in which a transfer target (target domain) robot efficiently learns a motion by using motion information obtained by a transfer source (source domain) robot.

Information (motion information) on which the robot is to act may be regarded as obtained, for example, by accumulating the correspondence between joint angles (joint values) and coordinates (end effectors) of the front end of the arm of the robot having certain physical properties (for example, the length of the arm or the number of joints). The use of such motion information enables the robot to act. Therefore, the physical properties play an important role in the transfer of the robot motion. However, it is often difficult to unify the physical properties of robots in various types of robots. Therefore, in the motion transfer, a process for adapting motion information obtained from another robot to the physical properties of the robot is important.

As in the above method, a technique of obtaining physical properties of the transfer target robot and then processing the action information on the transfer target robot to be suitable for the physical properties of the transfer source robot is generally employed. However, in this technique, some preliminary preparation such as measurement of physical properties of the transfer target robot is required. Further, if the physical properties of the transfer target robot change, or if an error occurs in the measurement thereof, it is difficult to accurately perform the action. In addition, humans and animals can learn actions without having to acquire information about their own physical properties in advance. Therefore, it seems a more realistic method to realize the motion transfer based on the experience of a real robot without obtaining in advance physical properties about the transfer target robot.

In this regard, non-patent document 1 proposes a technique for transferring an action sample of a transfer source robot to a transfer target robot by using a corresponding number of identical action samples obtained from both the transfer source robot and the transfer target robot even in a case where some physical properties of the transfer target robot are unknown. It should be understood that in this case, the action transfer is realized by fitting using matrix calculation.

Documents of the prior art

Non-patent document

Non-patent document 1: botend Bocsi and two others, [ Online ] aligned robot model-based transfer learning, [ Online ], 2013, [2014, 10, 6, retrieve ], URL: http:// www.cs.ubbcluj.ro/. bboti/pubs/ijcnn-2013. pdf

Disclosure of Invention

Technical problem

However, non-patent document 1 fails to describe details regarding the number of data samples required for motion transfer. It is considered that the same number of motion samples need to be acquired from both the transfer source robot and the transfer target robot. In particular, the techniques require that a large number of learning samples be acquired not only in the transfer source but also in the transfer target. This causes the following problems: the technique takes a significant amount of time to perform an action transfer; a large amount of labor time, cost, etc.; and it is difficult to quickly and accurately implement motion transfer.

In general, using a real robot as a transfer source robot on a simulator or in an experimental facility makes it relatively easy to acquire a large number of motion samples. However, it is assumed that the transfer target robot is a real robot provided for actual operation at home or the like. In this case, it is considered difficult to acquire the same number of learning samples as the transfer source robot. Therefore, in the transfer-target robot, there is a need for a method for realizing motion transfer with high efficiency without a step of acquiring a large number of learning samples.

The present invention has been made to solve the above-described problems, and an object of the present invention is to realize motion transfer from a transfer source robot by using a small number of learning samples acquired by a transfer target robot.

Solution to the problem

A motion transfer device as a first aspect of the present invention includes: a transfer source motion information acquisition unit configured to acquire first motion information including data indicating a plurality of motions of a transfer source robot; a transfer target action information acquisition unit configured to acquire second action information including data indicating a plurality of actions of the transfer target robot; and a correction unit configured to generate third action information for transferring the action of the transfer source robot to the transfer target robot by using the second action information and correcting the first action information according to a predetermined update formula, the second action information containing a smaller number of data than the first action information, the first to third action information including a set of data indicating one or more robot joint values and a set of data indicating coordinate values of a specific part of the robot, the correction unit being configured to: acquiring a joint value identical to the joint value included in the second motion information from the joint values included in the first motion information; calculating an error between coordinate values corresponding to the acquired joint values contained in the first motion information and coordinate values contained in the second motion information; selecting a maximum error, which is a maximum error among the calculated errors of the coordinate values; selecting a joint value corresponding to the maximum error included in the second motion information; and transmitting the maximum error to each of the coordinate values included in the first motion information by using the update formula, and generating the third motion information, wherein a parameter of the update formula includes an error between a joint value corresponding to each of the coordinate values included in the first motion information and a joint value corresponding to the maximum error.

A motion transfer apparatus as the second aspect of the present invention is the motion transfer apparatus described above, wherein the correction unit is preferably configured to repeatedly perform the transfer of the error using the update formula until the maximum error is smaller than a preset threshold.

A motion transfer device as a third aspect of the present invention is the motion transfer device described above, wherein the update formula is preferably expressed as x ═ x +2sgm (a, d) × Δ x, where x is the coordinate value, Δ x is the maximum error, sgm (a, d) is a Sigmoid function of a gain a and a variable d, and d is the error of the joint value as the parameter contained in the update formula.

A motion transfer apparatus as a fourth aspect of the present invention is preferably the above motion transfer apparatus, wherein the coordinate values contained in the first motion information are values obtained by dividing a plurality of coordinate values obtained by operating the transfer source robot by a maximum value among the plurality of coordinate values, the coordinate values contained in the second motion information are values obtained by dividing a plurality of coordinate values obtained by operating the transfer target robot by a maximum value among the plurality of coordinate values, and the parameter d of the update formula is a value obtained by dividing the error between the joint value corresponding to each of the coordinate values contained in the first motion information and the joint value corresponding to the maximum error by a maximum value among the errors.

A motion transfer method as a fifth aspect of the present invention includes: acquiring first motion information including data indicating a plurality of motions of a transfer source robot; acquiring second action information including data indicating a plurality of actions of the transfer target robot; generating third motion information for transferring the motion of the transfer source robot to the transfer target robot by using the second motion information and correcting the first motion information according to a predetermined update formula, wherein the number of data contained in the second motion information is smaller than the number of data contained in the first motion information, and the first to third motion information include a set of data indicating one or more robot joint values and a set of data indicating coordinate values of a robot specific part; acquiring a joint value identical to a joint value included in the second motion information from the joint values included in the first motion information; calculating an error between coordinate values corresponding to the acquired joint values contained in the first motion information and coordinate values contained in the second motion information; selecting a maximum error, which is a maximum error among the calculated errors of the coordinate values; selecting a joint value corresponding to the maximum error included in the second motion information; and transmitting the maximum error to each of the coordinate values included in the first motion information by using the update formula, and generating the third motion information, wherein a parameter of the update formula includes an error between a joint value corresponding to each of the coordinate values included in the first motion information and a joint value corresponding to the maximum error.

An operation transition program as a sixth aspect of the present invention causes a computer to execute: a process of acquiring first motion information including data indicating a plurality of motions of a transfer source robot; a process of acquiring second action information including data indicating a plurality of actions of the transfer target robot; and a process of generating third motion information for transferring the motion of the transfer source robot to the transfer target robot by correcting the first motion information according to a predetermined update formula using the second motion information, wherein the number of data contained in the second motion information is smaller than the number of data contained in the first motion information, and the first to third motion information include a set of data indicating one or more robot joint values and a set of data indicating coordinate values of a robot specific part, and the process of generating the third motion information includes: acquiring a joint value identical to the joint value included in the second motion information from the joint values included in the first motion information; calculating an error between coordinate values corresponding to the acquired joint values contained in the first motion information and coordinate values contained in the second motion information; selecting a maximum error, which is a maximum error among the calculated errors of the coordinate values; selecting a joint value corresponding to the maximum error included in the second motion information; and transmitting the maximum error to each of the coordinate values included in the first motion information by using the update formula, and generating the third motion information, wherein a parameter of the update formula includes an error between a joint value corresponding to each of the coordinate values included in the first motion information and a joint value corresponding to the maximum error.

The invention has the advantages of

According to the present invention, it is made possible to realize the action transfer from the transfer source robot by using a small number of learning samples acquired by the transfer target robot.

Drawings

Fig. 1 schematically shows a block diagram of the configuration of a motion transfer apparatus according to a first embodiment.

Fig. 2 shows a flowchart of the operation of the motion transfer apparatus according to the first embodiment. Which is an example of an implementation of the process shown in the flow chart.

Fig. 3 shows a diagram of an example of an operation algorithm of the motion transfer apparatus according to the first embodiment.

Fig. 4 shows a diagram of a schematic configuration of a two-degree-of-freedom robot arm.

Fig. 5 shows a diagram of a workspace on a two-dimensional plane (X-Y plane) of a two-degree-of-freedom robotic arm.

Fig. 6 shows a graphical representation of motion transfer errors for a two degree-of-freedom robotic arm.

Fig. 7 shows a diagram of the time required for the motion transfer of the two-degree-of-freedom robot arm.

Fig. 8 is a diagram showing a relationship between a transfer error and the length of the arm of the transfer target robot arm in the case where the motion transfer is performed by the motion transfer technique according to the first embodiment.

Fig. 9 is a diagram showing a relationship between a transfer error and a length of an arm of a transfer target robot arm in the case where the motion transfer is performed by the conventional motion transfer technique. This is a graph showing the results of comparative experiments.

Fig. 10 shows a diagram of control performance in the case of transition data generated using the motion transition technique according to the first embodiment.

Fig. 11 shows a graph of control performance in the case of using transfer data generated according to the conventional action transfer technique.

Fig. 12 shows a diagram of a schematic configuration of a three-degree-of-freedom robot arm.

Fig. 13 shows a graphical representation of the transfer error of a three-degree-of-freedom robotic arm.

Fig. 14 shows a diagram of the time required for the motion transfer of the three-degree-of-freedom robot arm.

Fig. 15 shows a diagram of the control performance of the three-degree-of-freedom robot arm.

Fig. 16 shows a diagram of a schematic configuration of a six degree-of-freedom robotic arm.

Fig. 17 shows a graphical representation of the transfer error of a six degree-of-freedom robotic arm.

Fig. 18 shows a diagram of the time required for the motion transfer of the six-degree-of-freedom robot arm.

Fig. 19 shows a diagram of the control performance of a six degree-of-freedom robotic arm.

FIG. 20 shows a diagram of the action transfer concept according to the present invention.

List of reference numerals

1-3 mechanical arm

10 holding unit

11 end effector

100 action transfer device

101 migration source action information acquisition unit

102 migration target action information acquisition unit

103 correction unit

104 output unit

First joint J11, J21, J31

Second joint J12, J22, J32

ARM1 first ARM

ARM2 second ARM

LIST wrist unit

Detailed Description

First, various conditions constituting the premise of the action transition according to the present invention will be described. The present invention proposes a technique for effecting motion transfer between robots under the following conditions (fig. 20). The following embodiments of the present invention are also realized under the following conditions. It should be understood that motion of the robot as used herein refers to moving a particular part of the robot, such as an end effector of an arm (the end of the arm corresponding to a human finger), to any desired position (coordinate value).

Condition 1: the number of joints of the transfer source robot and the number of joints of the transfer target robot are known and equal to each other.

Condition 2: physical properties other than the number of joints (e.g., length of the arm) are unknown.

Condition 3: a large number of motion samples may be taken from the transfer source robot. The transfer source robot is generally a robot existing on a simulator or in an experimental facility.

Condition 4: only a small number of motion samples can be taken from the transfer target robot. The transfer target robot is usually a real robot.

The motion transfer technique according to the present invention transfers the motion of the transfer source robot satisfying the above conditions to the transfer target robot. Action transfer as used herein refers to transferring a source data set D from_src(which is an experience database of actions created on the transfer source robot) creates transfer data D_trans(which is an empirical database of working spaces suitable for transferring the target robot). At this time, the data D is transferred_transIs created by: first, the transfer target robot is caused to perform a small number of actions to create a transfer target sample data set D_tgtThen using the transfer target sample data set D_tgtTo correct the transferred source data set D_src. Here, the migration target sample data set D_tgtIs transferred to the source data set D_srcSmall, yet transferred data D_transSize of and transfer source data set D_srcAre of the same or equal size.

Therefore, according to the present technology, learning (i.e., motion transfer) can be achieved using a smaller number of data sets than in the case where data sets are generated by causing the transfer target robot to actually perform a large number of motions. Moreover, the source data set D is transferred by copying as it is_srcGenerating transfer data D_transThe accuracy of the motion transfer is improved compared to the case of (2).

It should be understood that such as transferring the source data set D_srcMay simply be a set of joint values and coordinate values in pairs. In addition to simple databases (databases that record information about a simple set of pairs of joint values and coordinate values), empirical data sets (such as the transfer source data set D)_src) It may be a database obtained by compressing information such as Self-Organizing Maps (SOM) using competitive learning, Self-Organizing Incremental learning Neural networks (SOINN), which is a technique of online unsupervised learning capable of additional learning by improving the SOM. It should be understood that SOM and SOINN are described in the following

references

1 and 2, and thus detailed description thereof is omitted.

(reference 1) Kohonen T., "Ad hoc formation of topologically correct signatures", Biometrics, Vol.43, pp.59-69, 1982

(reference 2) Shen f., Hasegawa o., "a fast nearest neighbor classifier based on ad-hoc incremental neural networks", neural networks, 2008

A method for creating an experience database D for the actions of the robot will now be described. This method can be applied to transfer a source data set D_srcAnd transferring the target sample data set D_tgtIs created. Typically, when the physical properties of the robot are known, the end effector of the arm can be controlled to move to any desired position using well-known Inverse Kinematics (IK) techniques. However, the physical properties are not known in the current assumptions. Thus, in the technique according to the present invention, the experience-based IK is performed, thereby moving the end effector. Data to be acquired by a robot for performing experience-based IK and methods for performing IK using the data are described below.

First, a joint value set J, which is data indicating joint angles of all joints of the target robot, is generated. The set of joint values J may be expressed as follows. Here, "i" is an integer of 1 to n (where n is an integer of 2 or more), which is an index indicating the posture of the robot of which joint value has been sampled.

[ equation 1]

It should be understood that the component joint values of the joint value set J

[ formula 2]

Is a vector composed of components of the number of joints, and in the case of an m-axis robot (i.e., a robot having m joints), for example, the joint values are represented by m-dimensional vectors.

Further, the robot is caused to actually act using the joint values included in the joint value set J, and absolute coordinate values (hereinafter simply referred to as coordinate values) of the end effector of the arm are obtained by a plurality of sensor information (for example, visual information) about the robot. When the physical properties of the robot on a simulator or the like are known, the use of positive kinematics can reduce the time required to acquire the coordinate values. Then, a coordinate value set X, which is a data set collecting the calculated coordinate values, is generated. The set of coordinate values may be expressed as follows.

[ formula 3]

It should be understood that the component coordinate values as the set of coordinate values X

[ formula 4]

Is defined as a two-dimensional vector indicating coordinates on a plane or a three-dimensional vector indicating coordinates in a three-dimensional space. Since the coordinate values of the end effector are uniquely determined from the joint values of the robot, there is a functional relationship between the joint values and the coordinate values.

The set of joint values and the set of coordinate values obtained from these empirical or actual movements are represented by a data set D. The data set D is represented by the following formula.

[ formula 5]

First embodiment

In the action n transfer according to the first embodiment, a transfer target sample data set D as a small number of action samples in the transfer target robot is acquired_tgtAnd using a transfer target sample data set D_tgtSet of coordinate values of (X)_tgtAnd corrects the set of coordinate values X of the source data set of the transfer source robot according to the update formula created by the reference SOM_srcThereby generating transfer data D_trans。

The SOM algorithm transfer is an algorithm conceived from an update formula with reference to a conventionally known SOM (self-organizing map). SOM indicates the self-organizing map conceived by Kohonen et al as a technique to change the network structure (topology) according to the input. The SOM is an artificial neural network and has a mechanism to modify the topology of neurons according to inputs. The SOM learning technique is competitive learning and updates neurons near the input.

On the other hand, since the setting of the target problem in the conventional SOM is different from the setting of the target problem in the SOM algorithm transition, a new finding is added to the method of changing the update rate in the transition by the SOM algorithm. In conventional SOM, nodes in a space are updated and the distance between an input and a neighboring node of the input in the space is used to determine the update rate. SOM algorithm transfer, on the other hand, is based on the premise that there are two spaces, namely, joint value space (joint value set) and end effector coordinate space (coordinate value set). In the present technology, the coordinate value space is suitable as a space in which updating is performed depending on input data. However, as described above, since there is a constant constraint condition between the coordinate values and the joint values, it is important to consider the distance between the joint values in the joint value space. In view of this, in the present technology, the distance between the joint values in the joint value space can be reflected in the update of the coordinate value space, and thus the accuracy of the motion transition is improved.

Hereinafter, specificallyObtaining transfer data D Using SOM Algorithm transfer_transThe technique of (1). First, the configuration of the motion transfer device 100 according to the first embodiment of the present invention will be described. Fig. 1 schematically shows a block diagram of a motion transfer device 100 according to a first embodiment.

The operation transfer device 100 is an information processing device such as a server computer or a Personal Computer (PC). The exemplary motion transfer device 100 includes an arithmetic processing unit, a volatile or non-volatile storage device, and an input/output device. The arithmetic processing unit performs various controls based on a program stored in the storage device, thereby logically implementing each processing unit to be described later. The motion transfer apparatus 100 is not necessarily a physical single apparatus, but may be implemented using a plurality of information processing apparatuses by distributed processing. Also, the motion transfer apparatus 100 may be incorporated into, for example, a transfer target robot, or may be separate from the transfer target robot.

As shown in fig. 1, the action transfer device 100 has a transfer source action information acquisition unit 101, a transfer destination action information acquisition unit 102, a correction unit 103, and an output unit 104.

The transfer source motion information acquisition unit 101 acquires a transfer source data set D for causing the transfer source robot to act_src. Transferring a Source dataset D_srcSet of joint values J including end effector_srcAnd set of coordinate values X_src(i.e., D)_src＝<J_src,X_src>). In general, the transfer source data set D is generated by making the transfer source robot act several times on a simulator and in an experimental facility and acquiring pairs of end effectors and joint values corresponding to the end effectors_src. In this embodiment, it is assumed that the transfer source data set D is generated in advance_srcAnd the migration source motion information acquisition unit 101 acquires the migration source data set D from the input device or the storage device_src。

The transfer target action information acquisition unit 102 acquires a transfer target sample data set D obtained when causing the transfer target robot to perform a test action_tgt. Transfer target sample data set D_tgtSet of joint values J including end effector_tgtAnd set of coordinate values X_tgt(i.e., D)_tgt＝<J_tgt,X_tgt>). In general, the transfer target sample data set D may be generated by causing the transfer source robot to randomly act several times and acquiring coordinate values of a plurality of pairs of end effectors and joint values corresponding to the coordinate values_tgt. In this case, the source data set D is transferred_srcBy comparing the sizes of the target sample data sets D_tgtIs extremely small. In other words, the number of motion tests of the transfer target robot is smaller than the number of motion tests of the transfer source robot. In this embodiment, it is assumed that a transfer target sample data set D is generated in advance_tgtAnd the migration target action information acquisition unit 102 acquires the migration target sample data set D from the input device or the storage device_tgt。

The correction unit 103 performs a process for correcting the data set D of the transfer target sample by using the transfer target sample data set D_tgtFor transferring the source data set D_srcTo transfer data D_transA transformation process is performed. In this embodiment, the correction unit 103 performs the conversion process using SOM algorithm transfer.

The output unit 104 outputs the transfer data D generated externally by the correction unit 103_trans. The transfer target robot acquires transfer data D output by the output unit 104_transAnd acts based on the transfer data, whereby an action similar to that of the transfer source robot can be obtained.

Subsequently, the operation of the motion transfer apparatus 100 according to the first embodiment of the present invention will be described with reference to the flowchart of fig. 2. It should be understood that the algorithm shown in fig. 3 is an example of an implementation of the processing indicated by the above-described flow chart. In the algorithm of fig. 3, a denotes a certain set and num (a) denotes the number of elements contained in the set a. sgm (a, x) represents the Sigmoid function of gain a. a is_eIs a parameter for determining the gain a. e indicates a threshold value to which an error to be described later should be met.

Step S11: initial setting of parameter t

First, the correction unit 103 sets the parameter t as an index of repetitive operations as an initial valueThe initial value is "1" (fig. 3, first row). Hereinafter, it will satisfy, for example, 1. ltoreq. t.ltoreq.num (D)_tgt) The integer of (2) is set to t.

Step S12: data set normalization

The migration source action information acquisition unit 101 acquires the migration source data set D_src. Subsequently, the migration source action information acquisition unit 101 derives the source data set D from the migration source data set D_srcSet of coordinate values X contained in_srcIs selected to be the maximum value (max | X)_src| and set by grouping a set of coordinate values X_srcBy the selected maximum value, to define a new set of coordinate values X_src(FIG. 3, second row). Thus, the source data set D is transferred_srcCoordinate value set X contained therein_srcIs normalized.

The transfer target action information acquisition unit 102 acquires a transfer target sample data set D_tgt. Subsequently, the migration target action information acquisition unit 102 migrates the source data set D from the migration source data set D_tgtCoordinate value set X contained therein_tgtIs selected to be the maximum value (max | X)_tgt| and set by grouping a set of coordinate values X_tgtBy the selected maximum value, to define a new set of coordinate values X_tgt(FIG. 3, third row). Thus, the transfer target sample data set D_tgtCoordinate value set X contained therein_tgtIs normalized.

Step S13: initial setting of transfer data

The correction unit 103 generates transfer data D for transferring the movement of the transfer source robot to the transfer target robot_trans. At this time, as transfer data D_transThe transfer source data set D whose coordinate value set is normalized in step S12_srcIs set as the branch data D_trans(FIG. 3, fourth row).

Step S14: error of calculating coordinate value

For transfer target sample data set D_tgtSet of joint values J contained therein_tgtJoint value of

[ formula 6]

For the transfer data D, the correction unit 103_transSet of joint values J_transThe same joint value in

[ equation 7]

And (6) obtaining.

Subsequently, the correction unit 103 aims at

[ formula 8]

Coordinate values corresponding to each joint value

[ formula 9]

Retrieving transfer data D_transSet of coordinate values of (X)_trans. In addition, coordinate values corresponding to the same joint value

[ equation 10]

And coordinate value

[ formula 11]

The error Δ x between is calculated. This error calculation is for the sample data set corresponding to the branch targetD_tgtCoordinate value set X contained therein_tgtIs performed, whereby an error data set Δ X having the calculated error as its element can be calculated (fig. 3, fifth line).

Step S15: determining the maximum value of the error

The correction unit 103 determines the maximum value max (| Δ X |) (fig. 3, sixth line) among the errors contained in the error data set Δ X.

Step S16: maximum value of estimation error

The correction unit 103 determines whether the maximum value max (| Δ X |) is equal to or larger than a predetermined threshold e (fig. 3, seventh line).

Step S17: calculating the gain a

If the maximum value max (| Δ X |) is equal to or greater than the predetermined threshold value e, the correction unit 103 calculates the gain a of the Sigmoid function according to the following formula (fig. 3, eighth line).

[ formula 12]

Here, the correction unit 103 may optionally set and transfer the target sample data set D_tgtIs processed with the associated parameter a_e(FIG. 3, eighth line). In other words, the parameter a for setting the property of Sigmoid function used in the conversion process_eCan be set. By modifying the property of the Sigmoid function, the speed of the transformation process (learning process) can be adjusted. It should be understood that this parameter may be set as appropriate.

Step S18: calculating deviation of joint value

Subsequently, if the target sample data set D is transferred_tgtThe joint value corresponding to the maximum value max (| Δ X |) contained in (a) is

[ formula 13]

The correction unit 103 calculates the joint value

[ formula 14]

And transfer data D_transEach joint value of

[ formula 15]

Deviation d between_j

[ formula 16]

d_j＝||j_tgt-j_i||²…(16)

Further, each deviation is normalized by dividing it by the maximum value of the calculated deviation (fig. 3, ninth to eleventh rows).

[ formula 17]

Step S19: error propagation

The correction unit 103 transfers the error to the transfer data D using the Sigmoid function and according to the following update formula (18)_transThe coordinate values contained in (fig. 3, twelfth and thirteen rows).

[ formula 18]

x_i＝x_i+2sgm(a，-d_j)·max|ΔX|…(18)

Step S20: increment parameter t

After the error transfer, the correction unit 103 increments the parameter t (adds 1 to t, i.e., t ═ t +1) (fig. 3, sixteenth row).

Step S21: verifying the number of repetitive treatments

The correction unit 103 determines whether the parameter t is smaller than num (D)_tgt)。

If the parameter t is less than num (D)_tgt) Then go throughThe routine returns to step S14. Thus, using the transfer data D updated by error transfer is repeatedly performed_transError propagation (fig. 3, fourteenth and fifteenth rows and seventeenth and subsequent rows).

Step S22:

as the repetitive process continues (as the number of repetitive processes increases), the above-mentioned error becomes progressively smaller, so that the maximum value of the calculated error will also become progressively smaller. Further, due to repetitive processing, the maximum value of the error becomes smaller than the threshold value e, and then the normalized coordinate values are restored as the original coordinate values according to the following formula and the process is terminated.

[ formula 19]

X_trans＝max|X_tgt|·Xtrans…(19)

Even if the maximum value of the error is not less than the threshold e, the process terminates to become equal to num (D) at the parameter t_tgt) While avoiding divergence of the process.

In this embodiment, the calculation of the joint value space is performed before the calculation of the coordinate value space. This is because there may be a plurality of joint values corresponding to a certain coordinate value. In such a case, when the calculation of the coordinate value space is performed, it is impossible to acquire and recognize the joint value corresponding to the coordinate value.

It should be noted here that, in the above-described update formula (18), when the coordinate value space is updated based on the related-art SOM update formula, information on the joint value space is added. This is because, unlike the related art SOM, the present invention needs to process a plurality of spaces such as a joint value space and a coordinate value space.

In this embodiment, the motion transfer apparatus 100 transforms the data set of the transfer source robot using the SOM algorithm, thereby generating a data set optimized for the transfer target robot. In this conversion process, motion samples that are acquired in the transfer target robot and are less than the motion samples contained in the data set are used. Therefore, even in the case where a large number of motion samples are not obtained in the transfer target robot, the transfer target robot can acquire the motion of the transfer source robot at the transfer target. In other words, equivalent actions may be implemented and actions may be transferred.

Further, according to this embodiment, it is made possible to realize the action transfer with a smaller amount of calculation than in the related art. In particular, since only a smaller amount of learning samples are required in the transfer target robot than in the related art, it is made possible to significantly reduce the amount of calculation.

< experiment >

In order to verify the advantageous effects of the motion transfer technique according to the first embodiment, the present inventors conducted the following experiments with respect to three types of two-degree-of-freedom, three-degree-of-freedom, and six-degree-of-freedom.

< experiments Using two-degree-of-freedom robot >

In this experiment, a two-degree-of-freedom robot arm that operates in two dimensions was created on a simulator. Fig. 4 is a diagram showing a schematic configuration of the two-degree-of-freedom robot arm 1. The two-degree-of-freedom robot ARM1 includes a first joint J11, a second joint J12, a first ARM1, and a second ARM 2. The first joint J11 and the second joint J12 are configured to be rotatable about a Z axis as a rotation axis perpendicular to the X axis and the Y axis. The holding unit 10 for holding the robot ARM and the first ARM1 are connected to each other by a first joint J11. The first ARM1 and the second ARM2 are connected to each other by a second joint J12. The end effector 11 is attached to the tip of a second ARM 2. In other words, the first joint J11 corresponds to the shoulder joint, the first ARM1 corresponds to the upper ARM, the second joint J12 corresponds to the elbow joint, and the second ARM2 corresponds to the forearm.

In this experiment, the movement range of the first joint J11 and the second joint J11 was limited to the range from 0 ° to 180 °. The length of the first ARM1 of the transfer source robot is 0.300m, and the length of the second ARM2 of the transfer source robot is 0.250 m. The length of the first ARM1 of the transfer-target robot is 0.600m, and the length of the second ARM2 of the transfer-target robot is 0.200 m. Fig. 5 shows the workspace on a two-dimensional plane (X-Y plane) of a two-degree-of-freedom robotic arm.

In this experiment, the transfer source data set was configured by the joint value obtained when the joint of the transfer source robot arm moved by 1.80 ° and the coordinate value (coordinate value on the X-Y plane) of the end effector at the tip end of the robot arm at this time. In other words, the learning data obtained by one round of learning has a total of four dimensions including two-dimensional joint values and two-dimensional coordinate values. In this example, the amount of data is about 10,000.

The area target sample data set is configured by joint values randomly selected from the transfer source data set and coordinate values of the end effector of the transfer target robot arm obtained when the selected joint values are applied to the transfer target robot arm.

First, the relationship between the data of the transfer target sample data set and the transfer error is investigated. In the following, the transfer error is estimated using Root Mean Square Error (RMSE). The RMSE indicative of the transfer error is represented by the following equation.

[ formula 20]

Here, N is the number of learning data to transfer the source data set. X_transIs transferring the data set D_transThe coordinate values of the data of (1). X_ltgtIs a coordinate value of the transfer target sample data set having the same joint value of the same amount as the joint value of the transfer source data set.

Fig. 6 shows a motion transfer error of the two-degree-of-freedom robot arm. In this experiment, when no action transfer is performed, in other words, when the transfer source data set is applied to the transfer target robot arm as it is, the transfer error (RMSE) is 0.215 (m). As shown in fig. 6, it can be appreciated that the transfer error (RMSE) can be reduced as compared to typical techniques, here motion transfer techniques using LPAs.

Subsequently, the processing time required for the action transition was investigated. Hereinafter, here, the simulator is made to operate on a 3.50-GHz personal computer, and the relationship between the time required to complete the action transfer and the data amount of the transfer target sample data set is made. Fig. 7 shows the time required for the motion transfer of the two-degree-of-freedom robot arm. In a typical technique (LPA), the processing time monotonically increases with respect to the number of data of the transfer target sample data set. In contrast, according to the action transfer technique of the first embodiment, the processing time is generally about four seconds in the area where the number of data of the transfer target sample data set is 200 or more.

As described in the foregoing, according to the motion transition technique of the first embodiment, the transition error (RMSE) rapidly decreases with respect to the number of data of the transition target sample data set, and the processing time converges as the number of data of the transition target sample data set is about 200. In view of this, it can be understood that, in the case where the number of data of the transfer target sample data set is obtained of about 200, it is permissible to achieve the action transfer with sufficient accuracy.

Subsequently, the relationship between the length of the arm of the transfer target robot arm and the transfer error was investigated. Here, the length of the two arms varies from 0.100(m) to 10.1(m), and the pitch is 0.5 (m). Fig. 8 is a diagram showing a relationship between a transfer error and the length of the arm of the transfer target robot arm in the case where the motion transfer is performed by the motion transfer technique according to the first embodiment. Fig. 9 is a diagram showing a relationship between a transfer error and a length of an arm of a transfer target robot arm in the case where the motion transfer is performed by the conventional motion transfer technique. In fig. 8 and 9, the horizontal axis indicates the length of one ARM1 of the transfer-target robot ARM, and the vertical axis indicates the length of the other ARM2 of the transfer-target robot ARM. In fig. 8 and 9, denser patterns indicate smaller transfer errors (RMSE).

Subsequently, the control performance of the action transition was investigated. Here, target positions (coordinate values) of the transfer target robot arm are randomly generated, and joint values for moving the respective arms toward the generated positions are predicted using transfer data by Inverse Distance Weighting (IDW). Details regarding IDW are disclosed in the agenda two-dimensional interpolation function for irregularly spaced data, d.shepard 1968, the 23 rd ACM national conference, 1968. Further, each arm is moved using the predicted joint value, and the RMSE between the reached position and the target position is measured. In this example, a total of 200 steps of evaluation were performed.

Fig. 10 shows a diagram of control performance in the case of transition data generated using the motion transition technique according to the first embodiment. Fig. 11 shows a graph of control performance in the case of using transfer data generated according to the conventional action transfer technique. In fig. 10 and 11, the horizontal axis indicates the length of one ARM1 of the transfer-target robot ARM, and the vertical axis indicates the length of the other ARM2 of the transfer-target robot ARM. In fig. 10 and 11, denser patterns indicate smaller RMSEs. As shown in fig. 10 and 11, it can be understood that RMSE is reduced and control performance is improved according to the motion transfer technique according to the first embodiment.

< experiment Using three-degree-of-freedom robot >

Subsequently, a three-degree-of-freedom mechanical arm operating in three dimensions was created on the simulator for further experiments. Fig. 12 shows a diagram of a schematic configuration of the three-degree-of-freedom robot arm 2. The three-degree-of-freedom mechanical ARM2 comprises a first joint J21, a second joint J22, a first ARM ARM1 and a second ARM ARM 2. The first joint J21 is a two-degree-of-freedom joint and the second joint J22 is a one-degree-of-freedom joint. The holding unit 10 for holding the robot ARM and the first ARM1 are connected to each other by a first joint J21. The first ARM1 and the second ARM2 are connected to each other by a second joint J22. The end effector 11 is attached to the tip of a second ARM 2. In other words, the first joint J21 corresponds to the shoulder joint, the first ARM1 corresponds to the upper ARM, the second joint J22 corresponds to the elbow joint, and the second ARM2 corresponds to the forearm.

In this experiment, the length of the first ARM1 of the transfer source robot was 0.300m, and the length of the second ARM2 of the transfer source robot was 0.250 m. The length of the first ARM1 of the transfer-target robot is 0.600m, and the length of the second ARM2 of the transfer-target robot is 0.200 m. Also, the movable range of the first joint J21 in the yaw direction (in the horizontal plane) is-180 ° to 180 °, and the movable range in the pitch direction is 0 ° to 180 °. The movable range of the second joint J22 in the pitch direction (in the vertical plane) is 0 ° to 180 °.

In this experiment, the transfer source data set was configured by joint values obtained when the joints of the transfer source robot arm were moved randomly and coordinate values (coordinate values in the X-Y plane) of the end effector at the tip end of the robot arm at that time. In other words, the learning data obtained by one round of learning has a total of six dimensions including the three-dimensional joint value and the three-dimensional coordinate value. In this example, the amount of data is about 100,000.

First, the relationship between the data of the transfer target sample data set and the transfer error is investigated. Fig. 13 shows the transfer error of the three-degree-of-freedom robot arm. In this experiment, when no action transfer is performed, in other words, when the transfer source data set is applied to the transfer target robot arm as it is, the transfer error (RMSE) is 0.156 (m). As shown in fig. 13, it can be appreciated that the transfer error (RMSE) can be reduced as compared to a typical technique, here a motion transfer technique using LPA.

Subsequently, the processing time required for the action transition was investigated. Fig. 14 shows the time required for the motion transfer of the three-degree-of-freedom robot arm. In this example, it has been shown that the processing time of the action transfer technique according to the first embodiment becomes longer than in the typical technique (LPA) until the number of data reaches 5,000. This is considered to be caused by an increase in the dimensionality of the implementation of the data, which leads to an increase in the processing load. However, when the number of data is about 5,000, a sufficient number of data is obtained, and thus the processing time is considered to have converged.

Subsequently, in the same manner as in the case of the two-degree-of-freedom robot arm, the control performance of the motion transfer was investigated. In this example, a total of 200 steps of evaluation were performed. Fig. 15 shows the control performance of the three-degree-of-freedom robot arm. As shown in fig. 15, the technique according to the first embodiment exhibits good control performance until the number of data reaches 500, but the LPA and the technique according to the first embodiment exhibit substantially the same control performance when the number of data reaches 5,000.

< experiments Using six-degree-of-freedom robot >

Subsequently, a six-degree-of-freedom robotic arm operating in three dimensions was created on the simulator for further experiments. Fig. 16 shows a diagram of a schematic configuration of the six-degree-of-freedom robot arm 2. The six-degree-of-freedom robot ARM2 includes a first joint J31, a second joint J32, a first ARM1, a second ARM2, and a wrist unit LIST. The first joint J31 is a two-degree-of-freedom joint and the second joint J32 is a one-degree-of-freedom joint. The wrist unit LIST is three degrees of freedom and is configured as a typical universal joint. The holding unit 10 for holding the robot ARM and the first ARM1 are connected to each other by a first joint J31. The first ARM1 and the second ARM2 are connected to each other by a second joint J32. The end of the second ARM2 is connected to the end effector 11 by a wrist unit LIST. In other words, the first joint J21 corresponds to the shoulder joint, the first ARM1 corresponds to the upper ARM, the second joint J22 corresponds to the elbow joint, and the second ARM2 corresponds to the forearm.

In this experiment, the length of the first ARM1 of the transfer source robot was 0.300m, the length of the second ARM2 of the second robot was 0.250m, and the length of the wrist unit LIST of the transfer source robot was 0.15 m. The length of the first ARM1 of the transfer target robot is 0.600m, the length of the second ARM2 of the transfer target robot is 0.200m, and the length of the wrist unit LIST of the transfer target robot is 0.09 m. The movable range of the first joint J31 in the yaw direction (in the horizontal plane) is-85 ° to 85 °, and the movable range in the pitch direction is-175 ° to 115 °. The movable range of the second joint J32 in the pitch direction (in the vertical plane) is-155 ° to 0 °. The movable range of the wrist unit LIST in the roll direction is-125 ° to 125 °, the movable range in the pitch direction is-95 ° to 95 °, and the movable range in the yaw direction is-130 ° to 190 °.

In this experiment, the transfer source data set was configured by joint values obtained when the joints of the transfer source robot arm were moved randomly and coordinate values (coordinate values in the X-Y plane) of the end effector at the tip end of the robot arm at that time. In other words, the learning data obtained by one round of learning has nine dimensions including six-dimensional joint values and three-dimensional coordinate values in total. In this example, the amount of data is 500,000.

First, the relationship between the data of the transfer target sample data set and the transfer error is investigated. Fig. 17 shows a transfer error of a six-degree-of-freedom robot arm. Five motion transfers were performed to evaluate using the average of these transfer errors. In this experiment, when no action transfer is performed, in other words, when the transfer source data set is applied to the transfer target robot arm as it is, the transfer error (RMSE) is 0.206 (m). As shown in fig. 17, it can be appreciated that the transfer error (RMSE) can be reduced as compared to a typical technique, here a motion transfer technique using LPA. Also, in the case of transferring the number of data of the target sample data set, the unit error in the LPA increases without increasing the unit error in the action transfer technique according to the first embodiment, and the transfer error is allowed to decrease by about half of the original error.

Subsequently, the processing time required for the action transition was investigated. Fig. 18 shows the time required for the motion transfer of the six-degree-of-freedom robot arm. In this instance, the processing time in the motion transfer technique according to the first embodiment becomes longer than that in the typical technique (LPA). This is believed to be caused by a further increase in the dimensionality of the implementation of the data, which leads to a further increase in the processing load.

The control performance of the motion transfer was investigated in the same manner as in the case of the two-degree-of-freedom robot arm. In this example, a total of 200 steps of evaluation were performed. Fig. 19 shows the control performance of a six-degree-of-freedom robot arm. As shown in fig. 19, the technique according to the first embodiment exhibits good control performance regardless of the amount of data.

< other embodiment >

It will be understood that the present invention is not limited to the embodiments described above, and may be modified as appropriate without departing from the scope of the invention. For example, in the above embodiment, a Sigmoid function is employed for mapping the coordinate values of the transfer source and the coordinate values of the transfer destination. However, instead of the Sigmoid function, any function such as an arctan function (arctan) may be employed.

Further, the above-described embodiments disclose a technique of moving the translation information on the premise that there are two spaces (i.e., the joint value space and the coordinate value space). However, any suitable number N of spaces may be provided. In this case, the above-described update formula can be appropriately expanded according to the number of spaces. For example, coordinate values may be defined as a function of joint values and coordinate values by adding a space of sensor values. The sensor values may be, for example, data indicative of arm status (e.g., overlapping or bending arms). This enables the robot to act taking sensor values into account. Therefore, it is considered that the operation can be selected and generated according to the robot itself and the peripheral state of the robot.

Also, although the above embodiments have been described assuming that the present invention has a configuration mainly based on hardware, the present invention is not limited thereto. Any processing may be realized by causing a Central Processing Unit (CPU) to execute a computer program. In this case, the computer program may be stored using any type of non-transitory computer readable medium and provided to the computer. Non-transitory computer readable media include any type of tangible storage media. Examples of the non-transitory computer-readable storage medium include magnetic storage media (such as floppy disks, magnetic tapes, and hard disk drives), magneto-optical storage media (e.g., magneto-optical disks), CD-ROMs (read-only memories), CD-R, CD-R/ws, and semiconductor memory devices (such as mask ROMs, PROMs (programmable ROMs), EPROMs (erasable PROMs), flash ROMs, RAMs (random access memories), and the like). Further, the program may be provided to a computer using any type of transitory computer-readable medium. Examples of transitory computer readable media include electrical signals, optical signals, and electromagnetic waves. The transitory computer-readable medium may provide the program to the computer through a wired communication line such as an electric wire and an optical line or a wireless communication line.

The present application is based on and claims the priority of Japanese patent application number 2016-217353, filed 2016, 11, 7, 2016, the disclosure of which is incorporated herein by reference in its entirety.

Claims

1. A motion transfer apparatus, comprising:

a transfer source motion information acquisition unit configured to acquire first motion information including data indicating a plurality of motions of a transfer source robot;

a transfer target action information acquisition unit configured to acquire second action information including data indicating a plurality of actions of the transfer target robot; and

a correction unit configured to correct the first action information by using the second action information and according to a predetermined update formula based on a self-organizing map algorithm to generate third action information for transferring the action of the transfer source robot to the transfer target robot,

the number of data contained in the second motion information is smaller than the number of data contained in the first motion information,

the first to third motion information includes a set of data indicating one or more robot joint values and a set of data indicating coordinate values of a specific part of the robot,

the correction unit is configured to:

acquiring a joint value identical to the joint value included in the second motion information from the joint values included in the first motion information;

calculating an error between coordinate values corresponding to the acquired joint values contained in the first motion information and coordinate values contained in the second motion information;

selecting a maximum error, which is a maximum error among the calculated errors of the coordinate values;

selecting a joint value corresponding to the maximum error included in the second motion information; and

transmitting the maximum error to each of the coordinate values included in the first motion information by using the update formula, and generating the third motion information, wherein a parameter of the update formula includes an error between a joint value corresponding to each of the coordinate values included in the first motion information and a joint value corresponding to the maximum error,

wherein the correction unit is further configured to repeatedly perform the passing of the error using the update formula until the maximum error is less than a preset threshold, an

The second motion information is generated by causing the transfer source robot to act randomly a plurality of times and acquiring coordinate values of a plurality of pairs of end effectors and joint values corresponding to the coordinate values.

2. The motion transfer apparatus according to claim 1, wherein the correction unit is configured to repeatedly perform the transfer of the error using the update formula until the maximum error is less than a preset threshold.

3. The motion transfer apparatus according to claim 1 or 2, wherein the update formula is expressed as x ═ x +2sgm (a, d) × Δ x, where x is the coordinate value, Δ x is the maximum error, sgm (a, d) is a Sigmoid function of a gain a and a variable d, and d is the error as the parameter contained in the update formula.

4. The motion transfer apparatus according to claim 3, wherein the coordinate value included in the first motion information is a value obtained by dividing a plurality of coordinate values obtained by the transfer source robot operation by a maximum value among the plurality of coordinate values,

the coordinate values contained in the second action information are values obtained by dividing a plurality of coordinate values obtained by the transfer target robot operation by a maximum value among the plurality of coordinate values, and

the parameter d of the update formula is a value obtained by dividing the error between the joint value corresponding to each of the coordinate values contained in the first motion information and the joint value corresponding to the maximum error by the maximum value among the errors.

5. A method of action transfer, comprising:

acquiring first motion information including data indicating a plurality of motions of a transfer source robot;

acquiring second action information including data indicating a plurality of actions of the transfer target robot;

generating third motion information for transferring the motion of the transfer source robot to the transfer target robot by correcting the first motion information using the second motion information and according to a predetermined update formula based on a self-organizing map algorithm, wherein

The number of data contained in the second action information is smaller than the number of data contained in the first action information, an

The first to third motion information includes a set of data indicating one or more robot joint values and a set of data indicating coordinate values of a specific part of the robot;

wherein the passing of the error is repeatedly performed using the update formula until the maximum error is less than a preset threshold, an

6. A non-transitory computer-readable medium storing a motion transfer program, the motion transfer program causing a computer to execute:

a process of acquiring first motion information including data indicating a plurality of motions of a transfer source robot;

a process of acquiring second action information including data indicating a plurality of actions of the transfer target robot; and

a process of generating third action information for transferring the action of the transfer source robot to the transfer target robot by correcting the first action information using the second action information and according to a predetermined update formula based on a self-organizing map algorithm, wherein

The first to third motion information includes a set of data indicating one or more robot joint values and a set of data indicating coordinate values of a specific part of the robot, and

the process of generating the third action information includes: